README.md

metadata

title: Game 2048 Env
emoji: 🎮
colorFrom: yellow
colorTo: red
sdk: docker
app_port: 7860
tags:
  - openenv
  - reinforcement-learning
  - '2048'
  - game

2048 Game Environment for OpenEnv

A classic 2048 sliding tile puzzle environment for agentic RL training, built on the OpenEnv framework.

Overview

The agent controls a 4x4 grid where tiles slide and merge. Equal-valued tiles merge on collision, doubling their value. After each move a new tile (2 or 4) spawns. The episode ends when no valid moves remain. The goal is to reach the 2048 tile.

Actions: 0=up, 1=down, 2=left, 3=right

Reward: Score gained from tile merges on each step (e.g., merging two 16s yields reward 32)

Observation: Board state (4x4 grid), current score, legal actions, max tile, and text rendering

Quick Start

Install

pip install git+https://huggingface.co/spaces/openenv-community/game-2048-env

Use

from game_2048_env import Game2048Env, Game2048Action

env = Game2048Env(base_url="https://openenv-community-game-2048-env.hf.space")
env.connect()

result = env.reset()
print(result.observation.board_text)

# Take a step (0=up, 1=down, 2=left, 3=right)
result = env.step(Game2048Action(action=2))  # move left
print(f"Score: {result.observation.score}, Max tile: {result.observation.max_tile}")

env.close()

Run Locally

# Clone and install
git clone <this-repo>
cd game_2048_env
pip install -e .

# Start the server
uvicorn server.app:app --host 0.0.0.0 --port 8000

Environment Details

Property	Value
Action space	Discrete(4): up, down, left, right
Observation	4x4 int grid + score + legal actions
Reward	Merge score per step
Episode end	No valid moves remaining
Win condition	Creating a 2048 tile

For RL Training

This environment is compatible with:

• torchforge (PyTorch agentic RL)
• TRL (Hugging Face)
• Unsloth (efficient fine-tuning)
• Any framework supporting the OpenEnv API (reset(), step(), state())